Image Restoration Using Convolutional Auto-encoders with Symmetric Skip Connections

Image restoration, including image denoising, super resolution, inpainting,and so on, is a well-studied problem in computer vision and image processing,as well as a test bed for low-level image modeling algorithms. In this work, wepropose a very deep fully convolutional auto-encoder network for imagerestoration, which is a encoding-decoding framework with symmetricconvolutional-deconvolutional layers. In other words, the network is composedof multiple layers of convolution and de-convolution operators, learningend-to-end mappings from corrupted images to the original ones. Theconvolutional layers capture the abstraction of image contents whileeliminating corruptions. Deconvolutional layers have the capability to upsamplethe feature maps and recover the image details. To deal with the problem thatdeeper networks tend to be more difficult to train, we propose to symmetricallylink convolutional and deconvolutional layers with skip-layer connections, withwhich the training converges much faster and attains better results.